FiDeSR: High-Fidelity and Detail-Preserving One-Step Diffusion Super-Resolution

Aro Kim; Myeongjin Jang; Chaewon Moon; Youngjin Shin; Jinwoo Jeong; Sang-hyo Park

cs.CVMar 3, 2026

FiDeSR: High-Fidelity and Detail-Preserving One-Step Diffusion Super-Resolution

Authors: Aro Kim, Myeongjin Jang, Chaewon Moon, Youngjin Shin, Jinwoo Jeong, Sang-hyo Park

Organizations: Kyungpook National University · Korea Electronics Technology Institute

Abstract

Diffusion-based approaches have recently driven remarkable progress in real-world image super-resolution (SR). However, existing methods still struggle to simultaneously preserve fine details and ensure high-fidelity reconstruction, often resulting in suboptimal visual quality. In this paper, we propose FiDeSR, a high-fidelity and detail-preserving one-step diffusion super-resolution framework. During training, we introduce a detail-aware weighting strategy that adaptively emphasizes regions where the model exhibits higher prediction errors. During inference, low- and high-frequency adaptive enhancers further refine the reconstruction without requiring model retraining, enabling flexible enhancement control. To further improve the reconstruction accuracy, FiDeSR incorporates a residual-in-residual noise refinement, which corrects prediction errors in the diffusion noise and enhances fine detail recovery. FiDeSR achieves superior real-world SR performance compared to existing diffusion-based methods, producing outputs with both high perceptual quality and faithful content restoration. The source code will be released at: https://github.com/Ar0Kim/FiDeSR.

Figures & tables

Figure 2: Example failure cases of diffusion-based Real-ISR methods. (b) AddSR introduces structural distortion and low-frequency inconsistency. (c) OSEDiff loses high-frequency details, producing over-smoothed texture. (d) PiSA-SR generates excessive details. In contrast, (e) our method achieves both high fidelity and detail-preserving.

Figure 3: Overall framework of FiDeSR. (a) Training process: LQ image $x_{L}$ is encoded into a latent $z_{L}$ , and the Diffusion Network predicts a coarse residual $r$ , which is refined by LRRB to create refined latent $z_{r}$ . The training loss is guided by the DAW module to emphasize fine structural details, and the refined latent ${z_{r}}$ is decoded by the VAE Decoder. (b) Inference process: Following the single-step diffusion process, the LQ latent $z_{L}$ is processed by the Diffusion Network to predict a residual, which is then refined by the LRRB. Refined $z_{r}$ is enhanced by frequency components through the LFIM, and finally decoded by the VAE Decoder to produce the Super-Resolution (SR) image $x_{SR}$ .

Figure 4: Comparison of the pipeline among ESRGAN, PiSA-SR, and our FiDeSR. (a) ESRGAN employs RRDB-based restoration in pixel space. (b) PiSA-SR uses a one-step diffusion network to predict a global residual in the latent space. (c) Our FiDeSR introduces latent residual refinement block (LRRB) that progressively refines the residual.

Dataset	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	CLIPIQA $\uparrow$	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	FID $\downarrow$
DRealSR	StableSR-200s	27.93	0.7491	0.3306	0.2290	0.6197	6.3865	58.5502	0.5569	147.48
	SeeSR-50s	28.14	0.7713	0.3141	0.2298	0.6889	6.4633	64.7302	0.6016	146.98
	DiffBIR-50s	25.90	0.6245	0.4669	0.2298	0.6333	6.3275	66.1334	0.6221	180.34
	PASD-20s	29.17	0.7954	0.3176	0.2220	0.7071	7.7138	50.4422	0.4683	143.08
	AddSR-4s	26.62	0.7384	0.3695	0.2651	0.7121	7.8536	65.0625	0.5993	167.69
	OSEDiff-1s	27.92	0.7835	0.2967	0.2162	0.6956	6.4389	64.6979	0.5898	135.45

Table 1: Quantitative comparison with state-of-the-art DM-based SR methods on synthetic and real-world test datasets. The number of diffusion inference steps is indicated by ’s’. The best and second best results of each metric are highlighted in red and blue , respectively. The third best results are indicated by an underline .

Figure 5: Qualitative comparisons with state-of-the-art DM-based SR methods.

Model	CLIPIQA $\uparrow$	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$
w/o LRRB, DAW	0.6611	4.7381	67.6043	0.6237
w/o LRRB	0.6641	4.7129	67.6280	0.6236
w/o DAW	0.6626	4.7340	67.9515	0.6278
FiDeSR	0.6699	4.6300	68.2869	0.6285

Table 2: Ablation study on LRRB and DAW on the DIV2K dataset (evaluated before applying LFIM).

Method	DIV2K	DRealSR	RealSR	Average
baseline	0.1051	0.1049	0.1029	0.1049
LRRB	0.1038	0.1028	0.1045	0.1032
Improvement	1.24%	1.99%	1.62%	1.62%

Table 3: High-frequency noise prediction error comparison between baseline and LRRB models. Lower MSE indicates better performance.

Method	PSNR $\uparrow$	SSIM $\uparrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$
Baseline	26.2542	0.7498	69.3610	0.6580
HF-0.1	26.1295	0.7476	69.6450	0.6639
HF-0.2	25.9985	0.7452	69.8562	0.6692
HF-0.3	25.8618	0.7428	70.0120	0.6737
HF-0.4	25.7213	0.7403	70.1132	0.6774
HF-0.5	25.5778	0.7378	70.1617	0.6803

Table 4: Ablation of LFIM (HF/LF) modules with different injection on the RealSR dataset.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Ground-truth images used in the user study along with the voting results for their corresponding restored outputs.

Dataset	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	CLIPIQA $\uparrow$	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	FID $\downarrow$
DrealSR	Real-ESRGAN	28.61	0.8051	0.2819	0.2089	0.4519	6.6896	54.2678	0.4904	147.68
	BSRGAN	28.70	0.8028	0.2858	0.2144	0.5093	6.5387	57.1626	0.4844	155.59
	LDL	28.20	0.8124	0.2792	0.2127	0.4475	7.1360	53.9464	0.4894	155.53
	FiDeSR	28.90	0.7907	0.2836	0.2112	0.6974	6.2014	65.7820	0.6239	127.97
RealSR	Real-ESRGAN	25.68	0.7614	0.2709	0.2060	0.4485	5.7936	60.3674	0.5505	135.20
	BSRGAN	26.37	0.7651	0.2656	0.2124	0.5119	5.6361	63.2870	0.5420	141.30

Appendix

Table 5: Quantitative comparison with GAN-based Real-ISR Methods. Best results are highlighted in red .

Figure 7: Qualitative comparisons between FiDeSR and GAN-based Real-ISR methods.

Figure 8: Qualitative comparisons between FiDeSR and different diffusion-based methods on DIV2K dataset. FiDeSR effectively reconstructs fine details while preserving overall image fidelity.

Figure 9: Qualitative comparisons between FiDeSR and different diffusion-based methods on DRealSR and RealSR dataset. FiDeSR effectively reconstructs fine details while preserving overall image fidelity.

Figure 10: Spatial distribution of high-frequency noise prediction error improvement by LRRB. $\Delta E_{\mathrm{HF}}$ represents the difference between baseline and LRRB error magnitudes in the high-frequency domain (top 20% frequencies, $r_{c}=0.8$ ). Positive values (red) indicate regions where LRRB reduces prediction error, while negative values (blue) indicate regions where error increases. Color intensity represents the magnitude of change in log scale.

Figure 11: Visualization of the Detail-aware Weighting (DAW) module. Detail Map generated by spatial filters (Sobel, Laplacian, Variance), is element-wise multiplied by the Error Map to create the Difficulty Weight Map.

Figure 12: Qualitative ablation study illustrating the contributions of the LRRB and DAW modules. Even under challenging degradation conditions, the full FiDeSR model reconstructs finer details and preserves texture fidelity most effectively.

Metric	StableSR	DiffBIR	SeeSR	PASD	AddSR	SinSR	OSEDiff	PISA-SR	FiDeSR (Ours)
Inference Step	200	50	50	20	4	1	1	1	1
Inference Time (s)	7.52	2.04	3.30	2.10	0.76	0.097	0.087	0.057	0.078
#Params (B)	1.56	1.68	2.51	2.31	2.28	0.18	1.77	1.30	1.29

Appendix

Table 6: Comparison of inference steps, runtime, and model parameters among diffusion-based SR methods.

LoRA Rank	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	CLIPIQA $\uparrow$	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	FID $\downarrow$
4	26.5196	0.7577	0.2573	0.1967	0.6554	5.3488	68.4282	0.6442	112.9637
8	26.2542	0.7498	0.2604	0.1963	0.6764	5.3332	69.3610	0.6580	108.7867
16	26.4582	0.7563	0.2508	0.1920	0.6660	4.6858	68.0145	0.6525	108.6132

Appendix

Table 7: Ablation study of LoRA rank for FiDeSR on the RealSR dataset (evaluated before applying LFIM).

LF/HF Ratio	PSNR $\uparrow$	SSIM $\uparrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$
(0.2, 0.2)	26.0249	0.7457	0.6896	69.8245	0.6681
(0.4, 0.2)	26.0658	0.7464	0.6877	69.7645	0.6666
(0.4, 0.4)	25.8630	0.7428	0.6934	70.0136	0.6737
(0.6, 0.6)	25.6519	0.7391	0.6956	70.1429	0.6789

Appendix

Table 8: Ablation study on different LF/HF injection strength ratios of LFIM

(\texttt{lf\_alpha},\texttt{hf\_beta})

on the RealSR dataset.

Dataset	Method	PSNR $\uparrow$	LPIPS $\downarrow$	NIQE $\downarrow$	MANIQA $\uparrow$	FID $\downarrow$
DRealSR	TFDSR	27.88	0.3417	6.2667	0.6164	155.66
DRealSR	FiDeSR	28.90	0.2836	6.2014	0.6239	127.97

Appendix

Table 9: Comparison between FiDeSR and TFDSR on SR benchmarks.

Explore similar work

DreamSR: Towards Ultra-High-Resolution Image Super-Resolution via a Receptive-Field Enhanced Diffusion Transformer
May 15, 2026Qingji Dong, Hang Dong, Mingqin Chen +2Diffusion Models for Image Restoration Diffusion Model Guidance
ByteDance Inc.
GramSR: Visual Feature Conditioning for Diffusion-Based Super-Resolution
Apr 28, 2026Fabio D'Oronzio, Federico Putamorsi, Leonardo Zini +2Diffusion Models for Image Restoration Diffusion Model Fine-Tuning
University of Modena and Reggio Emilia, Italy
Fast Image Super-Resolution via Consistency Rectified Flow
May 12, 2026Jiaqi Xu, Wenbo Li, Haoze Sun +8Rectified Flow Consistency Distillation
The Chinese University of Hong Kong · Huawei Noah’s Ark Lab · HKUST (GZ) +1

May 15, 2026J/K move · Enter open · S save

DreamSR: Towards Ultra-High-Resolution Image Super-Resolution via a Receptive-Field Enhanced Diffusion Transformer

Qingji Dong, Hang Dong, Mingqin Chen, Rui Zhang+1

ByteDance Inc.

Diffusion Models for Image Restoration Diffusion Model Guidance Diffusion Transformer Image Super-Resolution

Open paper AskPDF arXiv

Large-scale pre-trained diffusion models have been extensively adopted for real-world image Super-Resolution because of their powerful generative priors through textual guidance. However, when super-resolving high-resolution images with patch-wise inference strategy, most existing diffusion-based SR methods tend to suffer from over-generation, due to the misalignment between the global prompt from LR image and the incomplete semantic information of local patches during each inference step. On the other hand, most existing methods also failed to generate detailed texture in local patches due to the overemphasis on global generation capabilities in network designs and training strategies. To address this issue, we present DreamSR, a novel SR model that suppresses local over-generation and improves fine-detail synthesis, thereby achieving visually faithful results with ultra-high-quality details. Specifically, we propose a dual-branch MM-ControlNet, where the ControlNet generates local textual feature with patch-level prompts while the pre-trained DiT provides global textual feature with global prompts, thereby mitigating over-generation and ensuring semantic consistency across patches. We also design a comprehensive training strategy with stage-specific data processing pipelines and a Receptive-Field Enhancement strategy, enhancing the model's capability to capture patch information and effectively restore local textures. Extensive experiments demonstrate that DreamSR outperforms state-of-the-art methods, providing high-quality SR results. Code and model are available at https://github.com/jerrydong0219/DreamSR.

Abstract

Explore similar work

DreamSR: Towards Ultra-High-Resolution Image Super-Resolution via a Receptive-Field Enhanced Diffusion Transformer

GramSR: Visual Feature Conditioning for Diffusion-Based Super-Resolution

Fast Image Super-Resolution via Consistency Rectified Flow